[Home]Export Modules/Export Ffmpeg

Showing revision 14
Difference (from revision 14 to revision 14) (minor diff, author diff)
(The revisions are identical or unavailable.)

Introduction

Based on mplayers DOCS/tech/libavc-option.txt

Modified for transcode.

Some (basic) options are supported through the transcode commandline and some through ffmpeg.cfg or ~/.ffmpeg.cfg. The format of ffmpeg.cfg is simple. It can consists of several sections, each prefix by a header [codec] which is the same string you give to transcode through -F. To see all available options, have a look at either export/ffmpeg_cfg.c:90ff or run transcode -y ffmpeg -F mpeg4 -q2

 --- ~/.ffmpeg.cfg ----
 [mpeg4]
 nr = 20
 mbd = 2
 trell = 1
 [mpeg1video]
 [mpeg2video]
 mbd = 2
 mv0 = 1
 trell = 1
 cbp = 1
 nr = 20
 vqcomp = 0.7
 vqblur = 0.3
 --- ~/.ffmpeg.cfg ----

Now to the options.

Description of what all those libavcodec options do ...

WARNING: I am no encoding expert so the recommendations might be bad ... if you find any errors, missing stuff, ... send a patch or cvs commit if you have an cvs account :)

Encoder options (lavcopts)

vqmin 1-31 (minimum quantizer) for pass1/2

 1 is not recommended (much larger file, little quality difference (if u are lucky) 
           and other weird things (if u are less lucky))
           weird things: msmpeg4, h263 will be very low quality
                         ratecontrol will be confused -> lower quality
                         some decoders will not be able to decode it
 2 is recommended for normal mpeg4/mpeg1video encoding (default)
 3 is recommended for h263(p)/msmpeg4
    the reason for 3 instead of 2 is that 2 could lead to overflows
           (this will be fixed for h263(p) by changing the quanizer per MB in     
           the future, but msmpeg4 doesnt support that so it cant be fixed for
    that)

 TRANSCODE: use --divx_quant min,max

vqscale 1-31 quantizer for constant quantizer / constant quality encoding

 1 is not recommended (much larger file, little quality difference and
        possible other weird things)
 lower means better quality but larger files
 see vqmin

 TRANSCODE: use --divx_quant min

vqmax 1-31 (maximum quantizer) for pass1/2

 31 default
 10-31 should be a sane range

 TRANSCODE: use --divx_quant min,max

mbqmin 1-31 (minimum macroblock quantizer) for pass1/2

 2 default

mbqmax 1-31 (maximum macroblock quantizer) for pass1/2

 31 default

vqdiff 1-31 (maximum quantizer difference between I or P frames) for pass1/2

 3 default

vmax_b_frames 0-4 (maximum number of B frames between non B frames)

 0 no b frames (default)
 0-2 is a sane range for mpeg4

vme 0-5 (motion estimation)

 0 none  (not recommanded, very lq)
 1 full  (not recommanded, too slow)
 2 log   (not recommanded, lq)
 3 phods (not recommanded, lq)
 4 EPZS  (default)
 5 X1    (experimantal, might change from time to time or be just broken)
        Note: 0-3 ignores the amount of bits spend currently, so dont complain
              about low quality

vhq (high quality mode)

 encode each MB as in all modes and choose the best (this is slow but
        better filesize/quality)
 enabled by default

v4mv

 allow 4 MV per MB (little difference in filesize/quality)
 disabled by default

keyint 0-300 (maximum interval between keyframes)

 keyframes are needed for seeking as seeking is only possible to a
        keyframe but they need more space than non-keyframes so larger numbers here
        mean slightly smaller files, but less precise seeking
 0 no keyframes 
 >300 is not recommended as the quality might be bad (depends upon
        decoder, encoder and luck)
        for strict mpeg1/2/4 compliance this would have to be <=132

 TRANSCODE: use -w X,X,keyint

vb_strategy 0-1 for pass 2

 0 allways use the max number of B frames (default)
 1 avoid B frames in high motion scenes (this will cause bitrate
          misprediction)

vpass

 1 first pass
        2 second pass
          (only need to specify if two-pass encoding is used)
        Tip: u can try to use constant quantizer mode for pass1 (vqscale=<quantizer>)
        for huffyuv:
         pass 1 saves statistics
                pass 2 encodes with a optimal huffman table based upon the pass 1 stats

 TRANSCODE: use -R (1|2)

vbitrate (kbits per second) for pass1/2

 800 is default
        (if value is bigger then 16000 it is interpreted as bit not kbit!)

 TRANSCODE: use -w bitrate

vratetol (filesize tolerance in kbit) for pass1/2

 this is just an approximation, the real difference can be much smaller
        or larger
 1000-100000 is a sane range
 8000 is default

vrc_maxrate (maximum bitrate in kbit/sec) for pass1/2

vrc_minrate (minimum bitrate in kbit/sec) for pass1/2

vrc_buf_size (buffer size in kbit) for pass1/2

 this is for stuff like VCD
 VCD:  FIXME
 SVCD: ...
 DVD:  1792
 Note: vratetol should not be too large during the 2.pass or there might 
              be problems if vrc_(min|max)rate is used

vb_qfactor (-31.0-31.0) for pass1/2

 1.25 is default
vi_qfactor (-31.0-31.0) for pass1/2
 0.8 is default
vb_qoffset (-31.0-31.0) for pass1/2
 1.25 is default
vi_qoffset (-31.0-31.0) for pass1/2
 0.0 is default
        if v{b|i}_qfactor > 0  
            I/B-Frame quantizer = P-Frame quantizer * v{b|i}_qfactor + v{b|i}_qoffset
        else    
            do normal ratecontrol (dont lock to next P frame quantizer) and
            set q= -q * v{b|i}_qfactor + v{b|i}_qoffset
 tip: to do constant quantizer encoding with different quantizers for
             I/P and B frames you can use:
             vqmin=<ip_quant>:vqmax=<ip_quant>:vb_qfactor=<b_quant/ip_quant>

vqblur (0.0-1.0) quantizer blur (pass1)

 0.0 qblur disabled
 0.5 is the default
 1.0 average the quantizer over all previous frames
 larger values will average the quantizer more over time so that it will
        be changed slower
vqblur (0.0-99.0) quantizer blur (pass2)
 gaussian blur (gaussian blur cant be done during pass 1 as the future quantizers arent known)
 0.5 is the default
 larger values will average the quantizer more over time so that it will
        be changed slower

vqcomp quantizer compression (for pass1/2)

 depends upon vrc_eq

vrc_eq the main ratecontrol equation (for pass1/2)

 1   constant bitrate
 tex   constant quality
 1+(tex/avgTex-1)*qComp approximately the equation of the old ratecontrol code
 tex^qComp  with qcomp 0.5 or something like that (default)

        infix operators: +,-,*,/,^
        variables:
  tex  texture complexity
  iTex,pTex intra, non intra texture complexity
  avgTex  average texture complexity
  avgIITex average intra texture complexity in I frames
  avgPITex average intra texture complexity in P frames
  avgPPTex average non intra texture complexity in P frames
  avgBPTex average non intra texture complexity in B frames
  mv  bits used for MVs
  fCode  maximum length of MV in log2 scale
  iCount  number of intra MBs / number of MBs
  var  spatial complexity
  mcVar  temporal complexity
  qComp  qcomp from the command line
  isI, isP, isB is 1 if picture type is I/P/B else 0
  Pi,E  see ur favorite math book

        functions: 
  max(a,b),min(a,b) maximum / minimum
  gt(a,b)   is 1 if a>b, 0 otherwise
  lt(a,b)   is 1 if a<b, 0 otherwise
  eq(a,b)   is 1 if a==b,0 otherwise
  sin,cos,tan,sinh,cosh,tanh,exp,log,abs

vrc_override user specified quality for specific parts (ending credits ...) (for pass1/2)

 <start-frame>,<end-frame>,<quality>[/<start-frame>,<end-frame>,<quality>[/...]]
 quality 2..31 -> quantizer
 quality -500..0 -> quality correcture in %

vrc_init_cplx (0-1000) initial complexity for pass1

vqsquish (0 or 1) for pass1/2 how to keep the quantizer between qmin & qmax

 0 use cliping
 1 use a nice differentiable function (default)

vlelim (-1000-1000) single coefficient elimination threshold for luminance

 0 disabled (default)
 -4 (JVT recommendation)
 negative values will allso consider the dc coefficient
 should be at least -4 or lower for encoding at quant=1

vcelim (-1000-1000) single coefficient elimination threshold for chrominance

 0 disabled (default)
 7 (JVT recommendation)
 negative values will allso consider the dc coefficient
 should be at least -4 or lower for encoding at quant=1

vstrict (-1,0,1) strict standard compliance

 0  (default)
        1  only recommended if you want to feed the output into the mpeg4 reference
           decoder
        -1 allows nonstandard YV12 huffyuv encoding (20% smaller files, but 
           cant be played back by the official huffyuv codec)

vdpart data partitioning

 adds 2 byte per video packet
 each videopacket will be encoded in 3 seperate partitions:
  1. MVs             (=movement)
  2. DC coefficients (=low res picture)
  3. AC coefficients (=details)
 the MV & DC are most important, loosing them looks far worse than 
 loosing the AC and the 1. & 2. partition (MV&DC) are far smaller than
 the 3. partition (AC) -> errors will hit the AC partition much more
 often than the MV&DC -> the picture will look better with partitioning
 than without, as without partitining an error will trash AC/DC/MV 
 equally
 improves error-resistance when transfering over unreliable channels (eg.
        streaming over the internet)

vpsize (0-10000) video packet size

 0 disabled (default)
 100-1000 good choice
 improves error-resistance (see vdpart for more info)

gray grayscale only encoding (a bit faster than with color ...)

nr (0-100000) (new option in cvs as of ~ sept 2005)

 In signal processing lingo, decimate is generally
 understood to mean filter and decimate to remove
 high frequency components that are not part of the 
 signal. This is what the nr option does (I think).
 Using values between 10-250 (this is the range I
 have tested) seem to be effective, with 20 as a
 good choice for a default for high quality source,
 and 120 as a good default for "average" quality
 source. The effect is similar to that of a good
 denoiser, but with very little overhead.
 It is a mystery to me why the scale for this option
 extends over 5 orders of magnitude.
 nr should ONLY EVER BE USED IN 2-PASS MODE!! In
 single-pass it is unpredictable and always BAD.

vfdct (0-99) dct algorithm

 0 automatically select a good one (default)
 1 fast integer 
        2 accurate integer
        3 mmx
        4 mlib

idct (0-99) idct algorithm

 0 automatically select a good one (default)
 1 jpeg reference integer
        2 simple
        3 simplemmx
        4 libmpeg2mmx (inaccurate, DONT USE for encoding with keyint >100)
        5 ps2
        6 mlib
        7 arm
        note: all these IDCTs do pass the IEEE1180 tests AFAIK

lumi_mask (0.0-1.0) luminance masking

 0.0 disabled (default)
        0.0-0.3 should be a sane range
        warning: be carefull, too large values can cause disasterous things
        warning2: large values might look good on some monitors but may look horrible
        on other monitors

dark_mask (0.0-1.0) darkness masking

 0.0 disabled (default)
        0.0-0.3 should be a sane range
        warning: be carefull, too large values can cause disasterous things
        warning2: large values might look good on some monitors but may look horrible
        on other monitors / TV / TFT

tcplx_mask (0.0-1.0) temporal complexity masking

 0.0 disabled (default)

scplx_mask (0.0-1.0) spatial complexity masking

 0.0 disabled (default)
        0.0-0.5 should be a sane range
        larger values help against blockiness, if no deblocking filter is used
        for decoding
        Tip: crop any black borders completly away as they will reduce the quality
        of the MBs there, this is true if scplx_mask isnt used at all too

naq normalize adaptive quantization (experimental)

 when using adaptive quantization (*_mask), the average per-MB quantizer
 may no longer match the requested frame-level quantizer. using naq will
 attempt to adjust the per-MB quantizers to maintain the proper average.

ildct use interlaced dct

format

 YV12 (default)
        422P (for huffyuv)

pred (for huffyuv)

 0 left prediction
        1 plane/gradient prediction
        2 median prediction

qpel use quarter pel motion compensation

 Tip: this seems only usefull for high bitrate encodings

precmp comparission function for motion estimation pre pass

cmp comparission function for full pel motion estimation

subcmp comparission function for sub pel motion estimation

mbcmp comparission function for macroblock decission

 0 SAD (sum of absolute differences) (default)
        1 SSE (sum of squared errors)
        2 SATD (sum of absolute hadamard transformed differences)
        3 DCT (sum of absolute dct transformed differences)
        4 PSNR (sum of the squared quantization errors) (dont use, low quality)
        5 BIT (number of bits needed for the block)
        6 RD (rate distoration optimal, slow)
        7 ZERO (0)
        +256 (use chroma too, doesnt work (correctly) with b frames currently)
        Tip: SAD is fast, SATD is good, SSE is something in between
        Tip2: when using SATD/DCT/BIT/RD for full pel search then a 
              larger diamond something like dia=2 or dia=4 might improve
              quality, but only for some videos
        Note: mbcmp is only used if hq mode is disabled

predia (-99 - 6) diamond type & size for motion estimation pre pass

dia (-99 - 6) diamond type & size for motion estimation

 ...
 -3 shape adaptive diamond with size 3
 -2 shape adaptive diamond with size 2
 -1 experimental
 1 normal size=1 diamond (default) =EPZS type diamond
                        0
                       000
                        0
 2 normal size=2 diamond
                        0
                       000
                      00000
                       000
                        0
        ...
        Tip: the shape adaptive stuff seems to be faster at the same quality
        Note: the sizes of the normal diamonds and shape adaptive ones dont
              have the same meaning

trell trellis quantization

 this will find the optimal encoding for each 8x8 block
 trellis quantization is quite simple a optimal quantization in the 
 PSNR vs bitrate sense (assuming that there would be no rounding errors introduced 
 by the IDCT, which is obviously not the case) it simply finds a block for the minimum of
 error + lambda*bits 
 lambda is a qp dependant constant
 bits is the amount of bits needed to encode the block
 error is simple the sum of squared errors of the quantization

last_pred (0-99) amount of motion predictors from the previous frame

 0 (default)
 a -> will use 2a+1 x 2a+1 MB square of MV predictors from the previous frame

preme (0-2) Motion estimation pre-pass

 0 disabled
 1 only after I frames (default)
 2 allways

subq (1-8) subpel refinement quality (for qpel)

 8 (default)
 Note: this has a significant effect on the speed

psnr will print the psnr for the whole video after encoding and store the per frame psnr

 in a file with name like "psnr_012345.log"

fps_code (1-8) will flag the framerate headers with the new framerate code

        0 - illegal
        1 - 24000.0/1001.0 (NTSC 3:2 pulldown converted FILM)
        2 - 24.0 (NATIVE FILM)
        3 - 25.0 (PAL/SECAM VIDEO / converted FILM)
        4 - 30000.0/1001.0 (NTSC VIDEO)
        5 - 30.0
        6 - 50.0 (PAL FIELD RATE)
        7 - 60000.0/1001.0 (NTSC FIELD RATE)
        8 - 60.0

TRANSCODE: The decoder options are currently unused in transcode

Decoder options (lavdopts)

ec error concealment
 1 use strong deblock filter for damaged MBs
        2 iterative MV search (slow)
        3 all (default)
        Note: just add the ones u want to enable

er error resilience

 0 disabled
        1 carefull (should work with broken encoders)
        2 normal (default) (works with compliant encoders)
        3 agressive (more checks but might cause problems even for valid bitstreams)
        4 very agressive

bug manual workaround encoder bugs (autodetection isnt foolproof for these)

 0 nothing
 1  autodetect bugs (default)
 2 for msmpeg4v3 some old lavc generated msmpeg4v3 files (no autodetect)
 4 for mpeg4 xvid interlacing bug     (autodetected if fourcc==XVIX)
 8 for mpeg4 UMP4                     (autodetected if fourcc==UMP4)
        16for mpeg4 padding bug
        32for mpeg4 illegal vlc bug          (autodetected per fourcc)
 64for mpeg4 XVID&DIVX qpel bug  (autodetected)
        Note: just add the ones u want to enable

gray grayscale only decoding (a bit faster than with color ...)

idct see lavcopts

        note: the decoding quality is highest if the same idct algorithm is used 
        for decoding as for encoding, this is often not the most accurate though

Notes

  1. Lavc will strictly follow the quantizer limits vqmin, vqmax, vqdiff even if it violates the bitrate / bitrate tolerance.
  2. Changing some options between pass1 & 2 can reduce the quality.

FAQ

Q: Why is the filesize much too small?

A: Try to increase vqmin=2 or 1 (be carefull with 1, it could cause strange things to happen).


Q: What provides better error recovery while keeping the filesize low?
Should I use data partitioning or increase the number of video packets?

A: Data partitioning is better in this case.

Glossary

MB
Macroblock (16x16 luminance and 8x8 chrominance samples)
MV
Motion vector
ME
Motion estimation
MC
Motion compensation
RC
Rate control
DCT
Discrete Cosine Transform
IDCT
Inverse Discrete Cosine Transform
DC
The coefficient of the constant term in the DCT (avg value of block)
JVT
Joint Video Team Standard -- http://www.itu.int/ITU-T/news/jvtpro.html
PSNR
peak signal to noise ratio

Examples

 mencoder foobar.avi -lavcopts vcodec=mpeg4:vhq:keyint=300:vqscale=2 -o new-foobar.avi

 mplayer  foobar.avi -lavdopts bug=1

Links

http://www.eecs.umich.edu/~amarathe/mpeg.html

http://www.cs.sfu.ca/undergrad/CourseMaterials/CMPT479/material/notes/Chap4/Chap4.2/Chap4.2.html
ftp://ftp.tek.com/mbd/manuals/video_audio/25W_11418_4.pdf


Written 2002 by Michael Niedermayer and reviewed by Felix Buenemann. Check the MPlayer documentation for contact-addresses.

Transcode Wiki | Export Modules | Recent Changes | Preferences
Password required to edit | View other revisions | View current revision
Edited February 23, 2008 1:05 am by Phil (diff)
Search: